Switch course ▾
Session 3. Checking AI Output for Errors and Bias
Objective. Check any AI-made resource with a four-point routine, record what you find in a checking log and decide whether to fix, replace or drop it.
Checking is the step that turns a risky draft into a usable resource, and it must be quick or it will not happen. Teachers who check well do not read everything with the same effort; they check the cheapest and most harmful problems first and match the effort to the stakes. This session gives you a short order of checks, shows what bias can look like in classroom materials in our schools, and helps you decide when a draft is not worth saving.
Key ideas
Why a fluent answer can be wrong
- These tools produce text that sounds likely, not text that has been checked. A wrong number or an invented source can sit inside a perfectly written paragraph.
- Typical problems for our classrooms: wrong arithmetic in answer keys, facts about Uzbekistan that are mixed up, content outside the state textbook, and references that do not exist.
- Quality can differ between languages: Uzbek and Russian output may contain odd terms or word-for-word translations. Read it aloud.
The four-point check
- Facts: recompute every number, check names, dates and places against the textbook or a second reliable source. Open any source the tool cites.
- Fit: does it match your objective, the textbook unit, the pupils' level, the language of instruction and the 45 minutes?
- Fairness and Language: who is missing or stereotyped, and does the wording read naturally and respectfully in Uzbek or Russian?
Recognise bias in classroom materials
- Look for patterns, not single words: only men in professional roles, only women at home, only city or wealthy settings, one region or one ethnic group as the default.
- Fix it with a new request ('use balanced roles, rural and city settings, names from different regions') and then read the new version again; fixes can add new problems.
- Your own judgement matters most: you know your pupils and what would make a child feel unseen.
Try it in your next lessons
- Keep a checking log for three AI-assisted resources: what was wrong, which check found it, how long checking took. Decision rule: if you cannot check it, you do not use it. Fallback: use the textbook version.
- Evidence: the log itself, with counts of problems by check (Facts, Fit, Fairness, Language).
- Evidence: one pupil-facing change, for example a corrected answer key, and a 5-minute note on whether the corrected resource worked in class.
The four checks at a glance
| Check | Ask | Typical find |
|---|---|---|
| Facts | Is it true? | Wrong sum, mixed-up place, invented source |
| Fit | Is it for my class? | Algebra in a fractions unit, too long |
| Fairness | Who is missing? | Stereotyped roles |
| Language | Does it read naturally? | Word-for-word translation |
Order of checks
A ten-minute order of checks
Start with what is cheapest to verify and costliest if wrong: every number and every answer key. Recompute them yourself, not by asking the tool whether they are right. Next check names, places and dates against the textbook, and open any source the tool cites; if you cannot find it, treat it as unverified.
Then check fit against your objective and the 45 minutes, read once for fairness, and read aloud in Uzbek or Russian for natural wording. Match effort to stakes: a test, an answer key or a notice to parents deserves a full pass; a brainstorm of ideas needs a quick read. If a draft needs more checking than writing would take, drop it.
What bias looks like in our materials
Many tools learn from text that comes mostly from other places, so their defaults may not match our classes: a family with a car and a tablet, a city flat, a snowy winter, names from one region. Look for who is missing: rural pupils, pupils whose home language is not the language of instruction, children from different regions, children with disabilities.
Also look at roles: who is the doctor, the driver, the teacher, the helper at home? One odd example is a slip; a pattern is bias. When you find one, give the tool a concrete correction, read the new version again, and make the final choice of names and settings yourself, because you know which examples would make your pupils feel seen.
Examples from the classroom
The source that was not there
Scene. Dilnoza teaches Grade 6 history in Samarkand. A tool wrote a ten-question quiz and ended it with a book title and page number. She spent three minutes looking in the school library list and the textbook and found nothing. She removed the line, checked each answer against the textbook, and replaced two questions that the textbook does not cover.
Insight. A source you cannot find is unverified, however precise it looks.
Everyone has a car
Scene. Malika teaches Grade 4 in a Russian-medium school in Tashkent. A word problem set assumed that every family drives to a supermarket and owns a tablet. She asked for rural and city settings, bazaar prices and walking distances, then read the new set again and changed one name that did not feel right for her class.
Insight. A bias fix is a new request plus a new read, and the final judgement is the teacher's.
A stiff notice
Scene. Aziz asked a tool to translate a notice about a school trip into Uzbek for parents. The text was correct but sounded like a legal form. He read it aloud, asked a colleague who writes well in Uzbek for a quick look, and rewrote three sentences in warmer words. Total time: ten minutes, which he judged better than writing it from zero.
Insight. The Language check is about tone and natural wording, not only correctness.
Common pitfalls
- Asking the tool 'is this correct?' instead of checking yourself.
- Checking the spelling and the layout but not the numbers.
- Fixing one stereotype and not reading the new version for new problems.
- Using a draft that took longer to check than to write.
Find the Planted Problems
- Alone (10 min): read the AI-style worksheet in the Activity Kit. Apply the four checks and circle every problem. Do not use a device: check with your own head and a pen.
- In pairs (10 min): compare, then fill the checking log: problem, which check found it, and what you would do (fix, replace or drop).
- In groups of four (10 min): rewrite the two worst items so they can be used, and write the follow-up prompt you would give a tool to fix the bias item.
- As a group (10 min): estimate the checking time. Is the saving worth it, or would the textbook version have been quicker? Agree one decision rule for your school.
Your facilitator will guide you through this activity in the session.
Questions to discuss
- Which problem was easiest to miss, and why?
- How long did checking take compared with writing the six items yourself?
- What is your personal limit: when do you stop and use the textbook version?
Reflect
- Which of the four checks do you usually skip when you are in a hurry?
- Have you ever used a resource, from any source, that later turned out to contain an error?
- Who at your school could be a second pair of eyes for you?
Try it tomorrow
- Pick one AI draft and recompute every number before reading anything else.
- Count roles and settings in one worksheet and note who is missing.
- Time yourself on one full check and write the minutes in your checking log.
Further reading and links
Check Facts, Fit, Fairness and Language every time; log what you find; if you cannot check it, do not use it.